6160Re: [decentralization] artificial immune system
- Oct 3, 2002I've attached the conference paper. It's contribution to the field
lies in the fact that Keith's implementation is distributed over a
small network. The fundamental ideas are found elsewhere (see the
So, an executive summary...
An artificial immume system is a model loosely based on the human
immune system. The objective is to rapidly identify invading
antigens (viruses, bacteria, or offending IP addresses) so that
other systems can halt the attack. In the body, this is done by
B-cells which bind specifically to a particular antigen. The trouble
is, there can only be about 10^8 (100 million!) different varieties
of B-cell in the body at any given time, but there are 10^12 to
10^16 possible antigens out there. The B-cells collectively form
a memory capable of storing 10^8 patterns, but must recognize 10^16
The body overcomes this problem by attempting to store the *right*
10^8 patterns, since there are probably not more than 10^8 antigens
it is likely to encounter. Besides, cells in the body itself
represent a large set of patterns, and it wouldn't do to have some
B-cells identify human cells as antigens.
So, the body manufactures immature B-cells in the thymus. Each new
B-cell recognizes some pattern in a space of 10^16 possibilities.
If the immature B-cell recognizes a human cell / pattern, a process
in the thymus kills it. If the B-cell does not recognize a human
pattern, it is let loose into the wild (your body).
Think of the B-cells as initially randomly distributed in this high-
dimensional (10^16 dimensions) space. Someone sneezes near you and
a set of antigens (viruses) lands in that space. One or two B-cells
happen to be near the antigen location and they bind to a few of the
virus particles. Most of the viruses get by and multiply like crazy,
causing you to get sick.
In the meantime, the B-cells that bound to viruses have signaled that
an attack is in progress. In response, the body begins manufacturing
copies of those B-cells by the truckload. Now it's a race between
two exponentially growing populations. Most of the time the B-cells
win and you get better.
After the fact, however, the distribution of B-cells in that high-
dimensional space is no longer random. In the vicinity of the
antigen pattern the B-cells are very dense. The next time that
particular antigen attacks (someone you gave your cold to sneezes),
that thick set of B-cells can grab just about every virus particle
that got in and kill it. The immune response is very quick and you
So after some time, your B-cells form a distributed memory of the
antigens you've encountered before. The distribution is thick in
the regions where likely antigens live and thin in regions where
they don't live.
That's how the body does it. Keith implemented his system with
mobile agents. These agents acted as carriers for digital B-cells,
moving patterns at random around the network. Antigens were
recognized in one part of the network, lots of digital B-cells were
generated in defense, and before long all the machines in the network
were able to recognize those antigens.
The trick lies in generating the patterns for the "B-cells" to bind
to. In the present case it should be quite simple: offending IP
addresses are the patterns. An attack on one part of the network
is detected and the appropriate "B-cells" then spread throughout the
network. The first attack succeeds to a degree but subsequent ones
fail. Moreover, the system is entirely de-centralized -- nowhere
can you find a single point of failure. This is achieved by simply
making each computer in the network act as its own "thymus".
It's fascinating work. I'm interested to hear what this group thinks
Lucas Gonze wrote:
> Rod Price wrote:
>>A co-worker of mine has implemented a de-centralized version of an
>>artificial immune system that would seem ideal for this application.
>>The system can recognize "self" and will flag "not-self." I know
>>this description is vague, but my co-worker isn't around right
>>now (9:00 pm) to help me out. For details on artificial immune
>>systems, look at http://www.cs.unm.edu/~forrest/papers.html,
>>particularly "Architecture for an Artificial Immune System" on
> Any chance of getting more of an executive summary, Rod? My first thought
> is that the idea of 'us' and 'them', as opposed to the 'me' and 'everybody
> else' that decentralized designs normally use, might be too slippery to
> work with.
> (posting that conference paper would be a good thing -- please do!)
> - Lucas
> To unsubscribe from this group, send an email to:
> Your use of Yahoo! Groups is subject to http://docs.yahoo.com/info/terms/
- << Previous post in topic Next post in topic >>